OpenAI previewed the GPT-5.6 family, Sol, Terra and Luna, on June 26 to roughly 20 pre-approved organizations, per VentureBeat, with the launch post noting the company had “previewed our plans and the models’ capabilities” to the U.S. government and was releasing “at the request” of the administration. That framing, buried under the product copy, is the news.
The gate exists because of a June 2 executive order from President Donald J. Trump directing federal agencies to construct a benchmarking process for frontier releases. The 30-day clock lands on July 2, which is also when broader access is expected “in the coming weeks,” alongside a Cerebras deployment of Sol promising up to 750 tokens per second for select customers. OpenAI’s own line: it’s “starting with a limited preview for a small group of trusted partners whose participation has been shared with the government.” The company doesn’t love this arrangement and says so plainly. “We don’t believe this kind of government access process should become the long-term default.”
The context around that sentence matters. TechCrunch reports that Anthropic released Fable 5 earlier this month, the administration ordered it to remove foreign-national access, and Anthropic pulled the model entirely; Claude Mythos 5 was “effectively banned” in the same window. GPT-5.6 is the first frontier launch to actually navigate the new process rather than collide with it.
The product surface is standard OpenAI tiering, with cosmological rebranding retiring the old “nano” and “mini” names. Sol runs $5 input and $30 output per million tokens; Terra, $2.50 and $15; Luna, $1 and $6. All three ship through the Codex API. TechCrunch calls Sol “slightly better” than Claude Mythos 5 on coding and competitive with the Mythos preview at roughly one-third of the output tokens, which is the more interesting number.
The preview system card discloses more than 700,000 A100-equivalent GPU hours spent on automated red-teaming for universal jailbreaks. Sol, Terra and Luna are rated High in Cybersecurity and Biological and Chemical risk under the Preparedness Framework; none reach High in AI Self-Improvement. In browser exploitation tests against Chromium and Firefox, Sol found bugs and exploitation primitives but didn’t autonomously chain them into a full exploit. Terminal-Bench 2.1 and GeneBench v1 scores round out the disclosures.
What’s structurally new isn’t the model. It’s the sequencing: an executive order, a pre-cleared partner list, a system card written for regulators as much as researchers, and a launch post that includes a polite objection to the very process it’s complying with. The lineage runs through the 2018 CFIUS expansion under FIRRMA, when Washington quietly rewired how frontier semiconductor and AI investments got reviewed. That precedent didn’t sunset. Neither, probably, will this one.
Sources
- https://openai.com/index/previewing-gpt-5-6-sol/
- https://deploymentsafety.openai.com/gpt-5-6-preview
- https://help.openai.com/en/articles/20001325-a-preview-of-gpt-56-sol-terra-and-luna
- https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/
- https://venturebeat.com/technology/openai-unveils-gpt-5-6-sol-terra-and-luna-models-but-only-accessible-to-limited-preview-partners-for-now-per-us-gov